Static and Dynamic Robustness in Emergency- Phase Communication - - PowerPoint PPT Presentation

Static and Dynamic Robustness in Emergency- Phase Communication Networks

Sean M. Fitzhugh1 and Carter T. Butts1,2

1Department of Sociology 2Institute of Mathematical Behavioral Sciences

University of California, Irvine MURI AHM May 25th, 2010

This material is based on research supported by the Office of Naval Research under award N00014-08-1-1015 as well as NSF award CMS-0624257

Outline

1.

Introduction: Network robustness and disaster response

2.

Methodology: How to measure network robustness

3.

Results and analysis: static case

4.

Dynamic robustness: Methods and results

5.

Concluding remarks

Network Robustness and Disaster Response

 Disaster response teams carry out complex tasks which

require extensive training and planning

 Typically operate in a volatile, chaotic environment  Perform tasks that require substantial coordination  Medical response/triage  Resource allocation  Search and rescue  Evacuation

Network Robustness and Disaster Response

 Certain types of network structure are conducive to

performing activities related to disaster response

 Locally centralized patterns of communication help

large groups of individuals carry out complex tasks (Bavelas 1973)

 To enhance efficiency, certain actors can function as

“information hubs”: may serve to coordinate actions

• f others

Network Robustness and Disaster Response

 Hub-dominated structure of observed WTC Radio

networks is potentially efficient, but this structure creates vulnerabilities

Network Robustness and Disaster Response

What happens if we eliminate the yellow node’s ties?

Network Robustness and Disaster Response

Note how many nodes have been isolated with the removal of just one individual.

Network Robustness and Disaster Response

What if we “remove” four more hubs?

Network Robustness and Disaster Response

Dozens more nodes have been isolated.

Network Robustness and Disaster Response

 This network’s information hubs are weak points

Network Robustness and Disaster Response

 Why is vulnerability problematic for these networks?  Without effective information transmission, tasks may

be carried out in an unstructured, counterproductive, or inefficient manner (Auf Der Heide 1989)

 Worse, some tasks may be overlooked altogether  Studying robustness patterns of communication

networks allows us to see who is important in holding the network together

 Actors with predetermined coordinative roles or

emergent coordinators?

Data: World Trade Center Radio

 Seventeen radio communications networks from the

World Trade Center disaster (Butts and Petrescu- Prahova, 2005)

 Fixed-channel radio communication: groups are

independent (no cross-channel radio communication), so we can think of them as separate

• rganizations

 Networks reconstructed from transcripts

 Transmission from actor i to actor j is coded as an (i,j) edge  Actors generally treat communication as dyadic

 Individual conversations dominate communication

Data: World Trade Center Radio

 Specialist networks: daily occupational routine

involves emergency response

 Lincoln Tunnel Police, Newark command, Newark Police,

Newark CPD, New Jersey Statewide Police Emergency Network (NJ SPEN1), NJ SPEN2, WTC Police, Port Authority Trans-Hudson (PATH) Police  Non-specialist networks: lack daily involvement in

emergency response, but were in some way involved with WTC response

 PATH radio communications, Newark operations terminals,

Newark maintenance, PATH control desk, WTC operations, WTC vertical transportation, Newark facility management, WTC maintenance electric

Data: World Trade Center Radio

 Each network has a number of actors in institutionalized

coordinative roles (ICR)

 Their formal role is to coordinate the actions of others in

the network

 Transcribed labels such as “command”, “desk”, “operator”,

“dispatch(er)”, “manager”, “control”, “base”

 Manage a variety of roles in these networks: assisting

searches for personnel, advising units on traffic/closures, coordinating equipment/EMT/personnel distribution, forwarding information

 Will ICRs operate in their formal, institutionalized roles or

will others adopt those roles?

How to Measure Network Robustness

 Test the robustness of a network by subjecting it to

various “attacks” (not literal attacks)

 Remove nodes from the network and see how well it

holds up

 Two basic sequences of node failure: random and

degree-targeted

 I also selectively target ICRs to assess their role in

holding the network together (leads me to use four total variations of sequential node failure)

 Remove nodes until none remain in the network

How to Measure Network Robustness

 Random failure: remove nodes at random

How to Measure Network Robustness

 Degree-targeted failure: remove nodes in sequential

• rder according to degree

How to Measure Network Robustness

 Random failure targeting ICRs: remove ICRs at random,

followed by random removal of remaining nodes

How to Measure Network Robustness

 Degree-targeted failure targeting ICRs: remove ICRs in

sequential order according to degree, followed by sequential removal of remaining nodes

How to Measure Network Robustness

 Connectivity:

 Who can reach whom?

 Isolate formation:

 Whose removal isolates others?

How to Measure Network Robustness

 Connectivity:

 Who can reach whom?

 Isolate formation:

 Whose removal isolates others?

How to Measure Network Robustness

 Connectivity:

 Who can reach whom?

 Isolate formation:

 Whose removal isolates others?

How to Measure Network Robustness

 Connectivity:

 Who can reach whom?

 Isolate formation:

 Whose removal isolates others?

Building Robustness Profiles

 We need a way to measure connectivity as a network

progressively degrades

 Robustness scores: measure of a network’s declining

connectivity as more and more of its nodes are removed

 Use simulation of node failure to obtain robustness

scores

 After up many iterations, simulation yields expected

mean connectivity as nodes are removed

 Let’s look at some examples for clarification…

Building Robustness Profiles

 Using either of the previous measures, plot the

robustness curve to monitor network connectivity as more nodes fail

Building Robustness Profiles

 Use multiple plots to compare robustness of different

series of node failures

The area between curves tells us how network robustness differs across attacks

Building Robustness Profiles

 Take the integral of the curve to obtain a robustness

score

Connectivity Random failure: 0.4287 Random failure of ICRs: 0.0397

Building Robustness Profiles

 Robust example:

Connectivity Random failure: 0.4159 Random failure of ICRs: 0.3579

Hypotheses

 With an understanding of how to measure network

robustness, we can test some hypotheses

 Hypothesis 1: Specialist and non-specialist networks will

be more robust to random failure than to random failure of ICRs

 Those with institutionalized roles will maintain those roles

during the disaster response

 Hypothesis 2: Specialist networks will be less robust to

loss of ICRs than non-specialist networks

 Trained for these types of tasks, specialists can consolidate

their coordination needs onto a smaller number of people

Hypotheses

 Hypothesis 3: Degree targeted failure and degree-

targeted failure of ICRs will produce similar robustness scores among specialist and non-specialist networks

 If ICRs occupy positions with the most ties, there should

be no difference between the two attacks

Comparing Robustness Profiles

 Calculate robustness scores for all varieties of attacks

(random, degree-targeted, and ICR-targeted) across measures of connectivity and isolate formation

 Use t-tests to compare scores across different

dimensions (ICR vs. non-ICR failures, specialist vs non- specialist networks)

Static Robustness: Results

 Static robustness examines the time-aggregated

networks

 Series of time-ordered communication events

collapsed into a single network

Static Robustness: Results

 Hypothesis 1: Specialist and non-specialist networks will be

more robust to random failure than to random failure of ICRs

 Hypothesis 2: Specialist networks will be less robust to loss

• f ICRs than non-specialist networks

 Specialist networks are significantly more robust to random

failure than to random failure of ICRs

 t=4.2877, p=.0026

 Among non-specialist networks, ICRs prove less crucial to

preserving connectivity

 t=1.9004, p=.0991

Static Robustness: Results

 Hypothesis 3: Degree targeted failure and degree-

targeted failure targeting ICRs will produce similar robustness scores among specialist and non-specialist networks

 Degree-targeted failure is significantly more damaging

than degree-targeted failure of ICRs in specialist networks

 t=-2.4815, p=.0380

 The difference between the two attacks is significant in

non-specialist networks

 t=-4.0548, p=.0048

Dynamic Robustness: Methodology

 Ordinal nature of transcripts allows us to explore

dynamic robustness

 Using the time-ordered sequence of communication to

measure forward connectedness

 How would network unfold if certain actors were

never present in the network?

Dynamic Robustness: Methodology

Can a message from A reach D? Time-aggregated network:

Dynamic Robustness: Methodology

Can a message from A reach D in the absence of C? Dynamic network:

Dynamic Robustness: Results

 Hypothesis 1: Specialist and non-specialist networks will be

more robust to random failure than to random failure of ICRs

 Hypothesis 2: Specialist networks will be less robust to loss

• f ICRs than non-specialist networks

 Difference between robustness scores of random failure and

random failure of ICRs remains significant for specialist networks

 t=3.5697,

p=0.0073  Random failure of ICRs remains not significantly more

damaging than random failure for non-specialists

 t=1.7971,

p=0.1154

Dynamic Robustness: Results

 Hypothesis 3: Degree targeted failure and degree-

targeted failure targeting ICRs will produce similar robustness scores among specialist and non-specialist networks

 Degree-targeted failure remains more damaging than

degree-targeted failure targeting ICRs

 t=-3.231,

p=0.005  Insignificant difference between specialist and non-

specialist robustness to degree-targeted failure

 t=0.778,

p=0.450

Results and Analysis: Recap

 What do these results tell us?

 Hypothesis 1: Rejected

 ICR failure is not significantly more damaging than random failure in

non-specialist networks (but ICRs still play an important role in specialist networks)

 Hypothesis 2: Supported

 ICRs play a more important role in coordinating specialist networks

than they do in non-specialist networks

 Hypothesis 3: Rejected

 Degree-targeted attack is more damaging than degree-targeted

attack on ICRs: it takes more than ICRs alone to hold together the network…

Isolate Formation: Results

 What can isolate formation tell us that connectivity cannot?  Measuring isolate formation tells us more about how these

attacks pull apart the networks

 Degree-targeted failure produces significantly more isolates

in specialist networks than it does in non-specialist networks

 t=-2.6515, p=.0237

 DT-ICR produces significantly more isolates in specialist

networks than it does in non-specialist networks

 t=-2.2608, p=.0441

Isolate Formation: Results

 What does this tell us that previous findings did not tell

us?

 When specialist networks lose their high-degree

actors (usually ICRs), many remaining actors become isolated

 Low degree actors tend to be tied exclusively to a single ICR

 Non-specialist networks have a higher level of

negotiation (more ties among those with relatively low numbers of ties)

Conclusions: What Have We Learned?

 Specialist networks are especially vulnerable to loss of ICRs

and subsequent node isolation

 Reliant on institutional features to build network structure  Non-specialist networks remain moderately more connected

following ICR loss

 Not as reliant on institutional roles to guide network structure  Relative lack of isolation suggests increased negotiation among non-

coordinators; likely have a more difficult time delegating emergency coordination tasks (have to figure out what to do and how to do it); confirmed in actual transcripts

Conclusions: Take-Home Points

 Organizational roles are key to predicting network

structure among specialists

 Non-specialists are less reliant on organizational

institutions to build their communication network

Future Directions

Static, time-aggregated robustness Dynamic robustness

 What’s next?  Resilience: How can the network actively respond to

damage and rebuild itself following personnel loss?

Thank you!

 Questions, comments, thoughts?

Dynamic Robustness: Methodology

 If dynamic gives a more precise result, why bother with

time-aggregated network?

 More precise for this exact ordering of ties

 Would network unfold exactly like this again? Can’t be sure

 Ties indicate open channel of communication

regardless of ordering of messages

 Illustrate opportunity structure for communication